Unsupervised Optical Flow Estimation Based on Multi-Scale Feature

Optical flow estimation is an essential task in the field of computer vision, aiming to capture the motion information of objects in image sequences over time. Optical flow estimation in traditional methods is typically modeled as an optimization problem, requiring the minimization of pixel-level brightness differences to compute optical flow, resulting in high computational complexity and demanding computational resources. With the advancement of deep learning technology, optical flow estimation has seen new breakthroughs. Deep learning models, usually based on Convolutional Neural Networks (CNNs), capture image features and motion patterns by learning from large-scale image data, thus achieving accurate optical flow estimation. Training deep learning models typically requires a large amount of labeled data, which can be difficult and expensive to obtain in the field of optical flow estimation. Therefore, on the one hand, we introduce an unsupervised optical flow estimation method based on photometric error design, which comprehensively considers pixel-level brightness differences, Structural Similarity Index (SSIM), and smoothness constraints. This method facilitates effective modeling and learning of optical flow estimation tasks in the absence of annotated data, enabling adaptation to various scenarios and data types, and exhibiting better generalization capability. On the other hand, we utilize short-term dense concatenate(STDC) modules to construct a feature encoding network, extracting multi-scale features with fewer parameters while reducing computational complexity.